snowmanliu / chipyard-gemmini-transformer

Chipyard Transformer Accelerator

This is a repo for constructing the working environment for developing Gemmini on a local machine. Also, if you want to reproduce the code in our paper, following steps here is the easiest method to go.

Environment (Workspace) Construction

The overall environmental setup is shown below. When completed, you’ll have git, VS Code, Docker and chipyard installed. Note that Gemmini is a directory in chipyard, so if you’ve installed chipyard, you will already install Gemmini.

Pull the docker images

docker pull chipyard/chipyard_transformer

Mount the git gemmini with docker container

https://github.com/BiEchi/chipyard

docker run -it --name=chipyard-transformer-software \
--mount type=bind,source="/Users/mac/Desktop/Research/chipyard"/gemmini,target=/root/chipyard/generators/gemmini chipyard/chipyard_transformer:software

docker run -it --name=chipyard-transformer-ws-sentence \
--mount type=bind,source="/home/user/gemmini"/software,target=/root/chipyard/generators/gemmini/software chipyard_wzs:v2

docker run -it --name=chipyard-transformer-software \
--mount type=bind,source="/mnt/d/Research/chipyard"/gemmini,target=/root/chipyard/generators/gemmini chipyard/chipyard_transformer:software

Use the software simulator

1. Spike

   spike --extension=gemmini <some/gemmini/baremetal/test>
   

2. Verilator

   cd ~/chipyard/sims/verilator # or "sims/vcs"
   make CONFIG=GemminiRocketConfig
   

Test Functionality

#!/bin/bash
# You're now at ~/chipyard/generators/gemmini/test
sbt "test:runMain modules.Launcher chiselPra_qly" # or other submodule you want to build

Project Process

TODO

Software (by C)

TODO	Executer	Day SPENT	PROCESS
math.h header file.	Hao BAI	1	DONE
softmax() implementation.	Liyang QIAN	3	PEND
normalize() implementation.	Liyang QIAN	3	DONE
WS/OS/CPU choices implementation.	Hao BAI	DONE	DONE
sin() for positional encoding.	Liyang QIAN	3	DONE
decoder() part.	Wentao YAO	3	DONE
segmented_linear_exponential.c implementation	Wentao YAO	3	DONE
segmented_linear_sqrroot.py implementation	Liyang QIAN	3	DONE

Hardware (by CHISEL)

TODO	Executer	Day SPENT	PROCESS
pipeline_divisor implementation	Wentao YAO	6	DONE
vector_ALU implementation	Hao BAI	3	DONE
pipeline_multiplicator implementation	Liyang QIAN	4	DONE
apply segmented_linear_exponential to hardware	Wentao YAO	7	ON
apply segmented_linear_sqrroot to hardware	Hao BAI	7	ON
Read the source code	Liyang QIAN	7	ON

FIXME

FIXME	Executer	Day SPENT	STATUS

Our Working Time

Time	Name
9-7-6	Liyang QIAN
10-7-6	Wentao YAO
9-5-6 / 9-9-5	Hao BAI

---

Working References

Layer Normalization

In our code, it shows as:

According to our analysis, we gain the formula $$ Layer\ Normalization(x_i)=\alpha\times\frac{x_i-\mu_L}{\sqrt{\sigma_L^2+\epsilon}}+\beta $$ Where $\alpha$, $\beta$ and $\epsilon$ are the super-parameters we deliver to the Layer Normalization Process; $\mu_L$ and $\sigma_L$ are the mean value and standard deviation of the LAYER, and $x_i$ is the value of the $i^{th}$ value of the current layer. To explain in graphs, we have:

Fix normalization()

Used segmented linear function to substitute the while statements.

Problem With Software-simulated Approach chipyard-verilator

Using spike software, we do not know how much time clocks it takes for a process. We could only simulate the result (structure).

We used chipyard-verilator to transform chisel into verilog, in order to simulate the time clocks. However, this approach is not exact. The estimated clocks should be 1,600, but the carried-out clocks is 1,300. The chipyard-verilator uses verilator to transform chisel into verilog, then run verilog on the simulated SoC-chip (hardware part). We then used build.sh to compile c files using risc-v (software part). Also, using this approach, the software compilation part is a bit too slow for PCs.

Switch chisel to verilog

This process can be found according to chipyard-document.

Read the source code of gemmini

Read all the source code files in /gemmini/src.

Learning Resources

CHISEL Learning Resources

CSE 293 Agile Hardware Design

Jupyter Notebook Script

Verliog Learing Web

https://hdlbits.01xz.net/wiki/Main_Page

Gemmini Learning

We’ve been looking into the overall architecture of Gemmini and carried out a good learning document in GemminiArchi.md.

We’ve also progressed on learning the Gemmini code and finished a documentation in GemminiCode.pdf.